drivers/net/virtio/virtio_rxtx_packed_avx.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2020 Intel Corporation
   3  */
   4
   5 #include <stdint.h>
   6 #include <stdio.h>
   7 #include <stdlib.h>
   8 #include <string.h>
   9 #include <errno.h>
  10
  11 #include <rte_net.h>
  12
  13 #include "virtio_logs.h"
  14 #include "virtio_ethdev.h"
  15 #include "virtio_pci.h"
  16 #include "virtqueue.h"
  17
  18 #define BYTE_SIZE 8
  19 /* flag bits offset in packed ring desc higher 64bits */
  20 #define FLAGS_BITS_OFFSET ((offsetof(struct vring_packed_desc, flags) - \
  21         offsetof(struct vring_packed_desc, len)) * BYTE_SIZE)
  22
  23 #define PACKED_FLAGS_MASK ((0ULL | VRING_PACKED_DESC_F_AVAIL_USED) << \
  24         FLAGS_BITS_OFFSET)
  25
  26 /* reference count offset in mbuf rearm data */
  27 #define REFCNT_BITS_OFFSET ((offsetof(struct rte_mbuf, refcnt) - \
  28         offsetof(struct rte_mbuf, rearm_data)) * BYTE_SIZE)
  29 /* segment number offset in mbuf rearm data */
  30 #define SEG_NUM_BITS_OFFSET ((offsetof(struct rte_mbuf, nb_segs) - \
  31         offsetof(struct rte_mbuf, rearm_data)) * BYTE_SIZE)
  32
  33 /* default rearm data */
  34 #define DEFAULT_REARM_DATA (1ULL << SEG_NUM_BITS_OFFSET | \
  35         1ULL << REFCNT_BITS_OFFSET)
  36
  37 /* id bits offset in packed ring desc higher 64bits */
  38 #define ID_BITS_OFFSET ((offsetof(struct vring_packed_desc, id) - \
  39         offsetof(struct vring_packed_desc, len)) * BYTE_SIZE)
  40
  41 /* net hdr short size mask */
  42 #define NET_HDR_MASK 0x3F
  43
  44 #define PACKED_BATCH_SIZE (RTE_CACHE_LINE_SIZE / \
  45         sizeof(struct vring_packed_desc))
  46 #define PACKED_BATCH_MASK (PACKED_BATCH_SIZE - 1)
  47
  48 #ifdef VIRTIO_GCC_UNROLL_PRAGMA
  49 #define virtio_for_each_try_unroll(iter, val, size) _Pragma("GCC unroll 4") \
  50         for (iter = val; iter < size; iter++)
  51 #endif
  52
  53 #ifdef VIRTIO_CLANG_UNROLL_PRAGMA
  54 #define virtio_for_each_try_unroll(iter, val, size) _Pragma("unroll 4") \
  55         for (iter = val; iter < size; iter++)
  56 #endif
  57
  58 #ifdef VIRTIO_ICC_UNROLL_PRAGMA
  59 #define virtio_for_each_try_unroll(iter, val, size) _Pragma("unroll (4)") \
  60         for (iter = val; iter < size; iter++)
  61 #endif
  62
  63 #ifndef virtio_for_each_try_unroll
  64 #define virtio_for_each_try_unroll(iter, val, num) \
  65         for (iter = val; iter < num; iter++)
  66 #endif
  67
  68 static inline void
  69 virtio_update_batch_stats(struct virtnet_stats *stats,
  70                           uint16_t pkt_len1,
  71                           uint16_t pkt_len2,
  72                           uint16_t pkt_len3,
  73                           uint16_t pkt_len4)
  74 {
  75         stats->bytes += pkt_len1;
  76         stats->bytes += pkt_len2;
  77         stats->bytes += pkt_len3;
  78         stats->bytes += pkt_len4;
  79 }
  80
  81 static inline int
  82 virtqueue_enqueue_batch_packed_vec(struct virtnet_tx *txvq,
  83                                    struct rte_mbuf **tx_pkts)
  84 {
  85         struct virtqueue *vq = txvq->vq;
  86         uint16_t head_size = vq->hw->vtnet_hdr_size;
  87         uint16_t idx = vq->vq_avail_idx;
  88         struct virtio_net_hdr *hdr;
  89         struct vq_desc_extra *dxp;
  90         uint16_t i, cmp;
  91
  92         if (vq->vq_avail_idx & PACKED_BATCH_MASK)
  93                 return -1;
  94
  95         if (unlikely((idx + PACKED_BATCH_SIZE) > vq->vq_nentries))
  96                 return -1;
  97
  98         /* Load four mbufs rearm data */
  99         RTE_BUILD_BUG_ON(REFCNT_BITS_OFFSET >= 64);
 100         RTE_BUILD_BUG_ON(SEG_NUM_BITS_OFFSET >= 64);
 101         __m256i mbufs = _mm256_set_epi64x(*tx_pkts[3]->rearm_data,
 102                                           *tx_pkts[2]->rearm_data,
 103                                           *tx_pkts[1]->rearm_data,
 104                                           *tx_pkts[0]->rearm_data);
 105
 106         /* refcnt=1 and nb_segs=1 */
 107         __m256i mbuf_ref = _mm256_set1_epi64x(DEFAULT_REARM_DATA);
 108         __m256i head_rooms = _mm256_set1_epi16(head_size);
 109
 110         /* Check refcnt and nb_segs */
 111         const __mmask16 mask = 0x6 | 0x6 << 4 | 0x6 << 8 | 0x6 << 12;
 112         cmp = _mm256_mask_cmpneq_epu16_mask(mask, mbufs, mbuf_ref);
 113         if (unlikely(cmp))
 114                 return -1;
 115
 116         /* Check headroom is enough */
 117         const __mmask16 data_mask = 0x1 | 0x1 << 4 | 0x1 << 8 | 0x1 << 12;
 118         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_off) !=
 119                 offsetof(struct rte_mbuf, rearm_data));
 120         cmp = _mm256_mask_cmplt_epu16_mask(data_mask, mbufs, head_rooms);
 121         if (unlikely(cmp))
 122                 return -1;
 123
 124         virtio_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
 125                 dxp = &vq->vq_descx[idx + i];
 126                 dxp->ndescs = 1;
 127                 dxp->cookie = tx_pkts[i];
 128         }
 129
 130         virtio_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
 131                 tx_pkts[i]->data_off -= head_size;
 132                 tx_pkts[i]->data_len += head_size;
 133         }
 134
 135         __m512i descs_base = _mm512_set_epi64(tx_pkts[3]->data_len,
 136                         VIRTIO_MBUF_ADDR(tx_pkts[3], vq),
 137                         tx_pkts[2]->data_len,
 138                         VIRTIO_MBUF_ADDR(tx_pkts[2], vq),
 139                         tx_pkts[1]->data_len,
 140                         VIRTIO_MBUF_ADDR(tx_pkts[1], vq),
 141                         tx_pkts[0]->data_len,
 142                         VIRTIO_MBUF_ADDR(tx_pkts[0], vq));
 143
 144         /* id offset and data offset */
 145         __m512i data_offsets = _mm512_set_epi64((uint64_t)3 << ID_BITS_OFFSET,
 146                                                 tx_pkts[3]->data_off,
 147                                                 (uint64_t)2 << ID_BITS_OFFSET,
 148                                                 tx_pkts[2]->data_off,
 149                                                 (uint64_t)1 << ID_BITS_OFFSET,
 150                                                 tx_pkts[1]->data_off,
 151                                                 0, tx_pkts[0]->data_off);
 152
 153         __m512i new_descs = _mm512_add_epi64(descs_base, data_offsets);
 154
 155         uint64_t flags_temp = (uint64_t)idx << ID_BITS_OFFSET |
 156                 (uint64_t)vq->vq_packed.cached_flags << FLAGS_BITS_OFFSET;
 157
 158         /* flags offset and guest virtual address offset */
 159         __m128i flag_offset = _mm_set_epi64x(flags_temp, 0);
 160         __m512i v_offset = _mm512_broadcast_i32x4(flag_offset);
 161         __m512i v_desc = _mm512_add_epi64(new_descs, v_offset);
 162
 163         if (!vq->hw->has_tx_offload) {
 164                 __m128i all_mask = _mm_set1_epi16(0xFFFF);
 165                 virtio_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
 166                         hdr = rte_pktmbuf_mtod_offset(tx_pkts[i],
 167                                         struct virtio_net_hdr *, -head_size);
 168                         __m128i v_hdr = _mm_loadu_si128((void *)hdr);
 169                         if (unlikely(_mm_mask_test_epi16_mask(NET_HDR_MASK,
 170                                                         v_hdr, all_mask))) {
 171                                 __m128i all_zero = _mm_setzero_si128();
 172                                 _mm_mask_storeu_epi16((void *)hdr,
 173                                                 NET_HDR_MASK, all_zero);
 174                         }
 175                 }
 176         } else {
 177                 virtio_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
 178                         hdr = rte_pktmbuf_mtod_offset(tx_pkts[i],
 179                                         struct virtio_net_hdr *, -head_size);
 180                         virtqueue_xmit_offload(hdr, tx_pkts[i], true);
 181                 }
 182         }
 183
 184         /* Enqueue Packet buffers */
 185         _mm512_storeu_si512((void *)&vq->vq_packed.ring.desc[idx], v_desc);
 186
 187         virtio_update_batch_stats(&txvq->stats, tx_pkts[0]->pkt_len,
 188                         tx_pkts[1]->pkt_len, tx_pkts[2]->pkt_len,
 189                         tx_pkts[3]->pkt_len);
 190
 191         vq->vq_avail_idx += PACKED_BATCH_SIZE;
 192         vq->vq_free_cnt -= PACKED_BATCH_SIZE;
 193
 194         if (vq->vq_avail_idx >= vq->vq_nentries) {
 195                 vq->vq_avail_idx -= vq->vq_nentries;
 196                 vq->vq_packed.cached_flags ^=
 197                         VRING_PACKED_DESC_F_AVAIL_USED;
 198         }
 199
 200         return 0;
 201 }
 202
 203 static inline int
 204 virtqueue_enqueue_single_packed_vec(struct virtnet_tx *txvq,
 205                                     struct rte_mbuf *txm)
 206 {
 207         struct virtqueue *vq = txvq->vq;
 208         struct virtio_hw *hw = vq->hw;
 209         uint16_t hdr_size = hw->vtnet_hdr_size;
 210         uint16_t slots, can_push;
 211         int16_t need;
 212
 213         /* How many main ring entries are needed to this Tx?
 214          * any_layout => number of segments
 215          * default    => number of segments + 1
 216          */
 217         can_push = rte_mbuf_refcnt_read(txm) == 1 &&
 218                    RTE_MBUF_DIRECT(txm) &&
 219                    txm->nb_segs == 1 &&
 220                    rte_pktmbuf_headroom(txm) >= hdr_size;
 221
 222         slots = txm->nb_segs + !can_push;
 223         need = slots - vq->vq_free_cnt;
 224
 225         /* Positive value indicates it need free vring descriptors */
 226         if (unlikely(need > 0)) {
 227                 virtio_xmit_cleanup_inorder_packed(vq, need);
 228                 need = slots - vq->vq_free_cnt;
 229                 if (unlikely(need > 0)) {
 230                         PMD_TX_LOG(ERR,
 231                                    "No free tx descriptors to transmit");
 232                         return -1;
 233                 }
 234         }
 235
 236         /* Enqueue Packet buffers */
 237         virtqueue_enqueue_xmit_packed(txvq, txm, slots, can_push, 1);
 238
 239         txvq->stats.bytes += txm->pkt_len;
 240         return 0;
 241 }
 242
 243 uint16_t
 244 virtio_xmit_pkts_packed_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
 245                         uint16_t nb_pkts)
 246 {
 247         struct virtnet_tx *txvq = tx_queue;
 248         struct virtqueue *vq = txvq->vq;
 249         struct virtio_hw *hw = vq->hw;
 250         uint16_t nb_tx = 0;
 251         uint16_t remained;
 252
 253         if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
 254                 return nb_tx;
 255
 256         if (unlikely(nb_pkts < 1))
 257                 return nb_pkts;
 258
 259         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
 260
 261         if (vq->vq_free_cnt <= vq->vq_nentries - vq->vq_free_thresh)
 262                 virtio_xmit_cleanup_inorder_packed(vq, vq->vq_free_thresh);
 263
 264         remained = RTE_MIN(nb_pkts, vq->vq_free_cnt);
 265
 266         while (remained) {
 267                 if (remained >= PACKED_BATCH_SIZE) {
 268                         if (!virtqueue_enqueue_batch_packed_vec(txvq,
 269                                                 &tx_pkts[nb_tx])) {
 270                                 nb_tx += PACKED_BATCH_SIZE;
 271                                 remained -= PACKED_BATCH_SIZE;
 272                                 continue;
 273                         }
 274                 }
 275                 if (!virtqueue_enqueue_single_packed_vec(txvq,
 276                                         tx_pkts[nb_tx])) {
 277                         nb_tx++;
 278                         remained--;
 279                         continue;
 280                 }
 281                 break;
 282         };
 283
 284         txvq->stats.packets += nb_tx;
 285
 286         if (likely(nb_tx)) {
 287                 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
 288                         virtqueue_notify(vq);
 289                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
 290                 }
 291         }
 292
 293         return nb_tx;
 294 }
 295
 296 /* Optionally fill offload information in structure */
 297 static inline int
 298 virtio_vec_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
 299 {
 300         struct rte_net_hdr_lens hdr_lens;
 301         uint32_t hdrlen, ptype;
 302         int l4_supported = 0;
 303
 304         /* nothing to do */
 305         if (hdr->flags == 0)
 306                 return 0;
 307
 308         /* GSO not support in vec path, skip check */
 309         m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
 310
 311         ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
 312         m->packet_type = ptype;
 313         if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
 314             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
 315             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
 316                 l4_supported = 1;
 317
 318         if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 319                 hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
 320                 if (hdr->csum_start <= hdrlen && l4_supported) {
 321                         m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
 322                 } else {
 323                         /* Unknown proto or tunnel, do sw cksum. We can assume
 324                          * the cksum field is in the first segment since the
 325                          * buffers we provided to the host are large enough.
 326                          * In case of SCTP, this will be wrong since it's a CRC
 327                          * but there's nothing we can do.
 328                          */
 329                         uint16_t csum = 0, off;
 330
 331                         rte_raw_cksum_mbuf(m, hdr->csum_start,
 332                                 rte_pktmbuf_pkt_len(m) - hdr->csum_start,
 333                                 &csum);
 334                         if (likely(csum != 0xffff))
 335                                 csum = ~csum;
 336                         off = hdr->csum_offset + hdr->csum_start;
 337                         if (rte_pktmbuf_data_len(m) >= off + 1)
 338                                 *rte_pktmbuf_mtod_offset(m, uint16_t *,
 339                                         off) = csum;
 340                 }
 341         } else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
 342                 m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
 343         }
 344
 345         return 0;
 346 }
 347
 348 static inline uint16_t
 349 virtqueue_dequeue_batch_packed_vec(struct virtnet_rx *rxvq,
 350                                    struct rte_mbuf **rx_pkts)
 351 {
 352         struct virtqueue *vq = rxvq->vq;
 353         struct virtio_hw *hw = vq->hw;
 354         uint16_t hdr_size = hw->vtnet_hdr_size;
 355         uint64_t addrs[PACKED_BATCH_SIZE];
 356         uint16_t id = vq->vq_used_cons_idx;
 357         uint8_t desc_stats;
 358         uint16_t i;
 359         void *desc_addr;
 360
 361         if (id & PACKED_BATCH_MASK)
 362                 return -1;
 363
 364         if (unlikely((id + PACKED_BATCH_SIZE) > vq->vq_nentries))
 365                 return -1;
 366
 367         /* only care avail/used bits */
 368 #if defined(RTE_ARCH_I686)
 369         __m512i v_mask = _mm512_set4_epi64(PACKED_FLAGS_MASK, 0x0,
 370                                            PACKED_FLAGS_MASK, 0x0);
 371 #else
 372         __m512i v_mask = _mm512_maskz_set1_epi64(0xaa, PACKED_FLAGS_MASK);
 373 #endif
 374         desc_addr = &vq->vq_packed.ring.desc[id];
 375
 376         __m512i v_desc = _mm512_loadu_si512(desc_addr);
 377         __m512i v_flag = _mm512_and_epi64(v_desc, v_mask);
 378
 379         __m512i v_used_flag = _mm512_setzero_si512();
 380         if (vq->vq_packed.used_wrap_counter)
 381 #if defined(RTE_ARCH_I686)
 382                 v_used_flag = _mm512_set4_epi64(PACKED_FLAGS_MASK, 0x0,
 383                                                 PACKED_FLAGS_MASK, 0x0);
 384 #else
 385                 v_used_flag = _mm512_maskz_set1_epi64(0xaa, PACKED_FLAGS_MASK);
 386 #endif
 387
 388         /* Check all descs are used */
 389         desc_stats = _mm512_cmpneq_epu64_mask(v_flag, v_used_flag);
 390         if (desc_stats)
 391                 return -1;
 392
 393         virtio_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
 394                 rx_pkts[i] = (struct rte_mbuf *)vq->vq_descx[id + i].cookie;
 395                 rte_packet_prefetch(rte_pktmbuf_mtod(rx_pkts[i], void *));
 396
 397                 addrs[i] = (uintptr_t)rx_pkts[i]->rx_descriptor_fields1;
 398         }
 399
 400         /*
 401          * load len from desc, store into mbuf pkt_len and data_len
 402          * len limiated by l6bit buf_len, pkt_len[16:31] can be ignored
 403          */
 404         const __mmask16 mask = 0x6 | 0x6 << 4 | 0x6 << 8 | 0x6 << 12;
 405         __m512i values = _mm512_maskz_shuffle_epi32(mask, v_desc, 0xAA);
 406
 407         /* reduce hdr_len from pkt_len and data_len */
 408         __m512i mbuf_len_offset = _mm512_maskz_set1_epi32(mask,
 409                         (uint32_t)-hdr_size);
 410
 411         __m512i v_value = _mm512_add_epi32(values, mbuf_len_offset);
 412
 413         /* assert offset of data_len */
 414         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) !=
 415                 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
 416
 417         __m512i v_index = _mm512_set_epi64(addrs[3] + 8, addrs[3],
 418                                            addrs[2] + 8, addrs[2],
 419                                            addrs[1] + 8, addrs[1],
 420                                            addrs[0] + 8, addrs[0]);
 421         /* batch store into mbufs */
 422         _mm512_i64scatter_epi64(0, v_index, v_value, 1);
 423
 424         if (hw->has_rx_offload) {
 425                 virtio_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
 426                         char *addr = (char *)rx_pkts[i]->buf_addr +
 427                                 RTE_PKTMBUF_HEADROOM - hdr_size;
 428                         virtio_vec_rx_offload(rx_pkts[i],
 429                                         (struct virtio_net_hdr *)addr);
 430                 }
 431         }
 432
 433         virtio_update_batch_stats(&rxvq->stats, rx_pkts[0]->pkt_len,
 434                         rx_pkts[1]->pkt_len, rx_pkts[2]->pkt_len,
 435                         rx_pkts[3]->pkt_len);
 436
 437         vq->vq_free_cnt += PACKED_BATCH_SIZE;
 438
 439         vq->vq_used_cons_idx += PACKED_BATCH_SIZE;
 440         if (vq->vq_used_cons_idx >= vq->vq_nentries) {
 441                 vq->vq_used_cons_idx -= vq->vq_nentries;
 442                 vq->vq_packed.used_wrap_counter ^= 1;
 443         }
 444
 445         return 0;
 446 }
 447
 448 static uint16_t
 449 virtqueue_dequeue_single_packed_vec(struct virtnet_rx *rxvq,
 450                                     struct rte_mbuf **rx_pkts)
 451 {
 452         uint16_t used_idx, id;
 453         uint32_t len;
 454         struct virtqueue *vq = rxvq->vq;
 455         struct virtio_hw *hw = vq->hw;
 456         uint32_t hdr_size = hw->vtnet_hdr_size;
 457         struct virtio_net_hdr *hdr;
 458         struct vring_packed_desc *desc;
 459         struct rte_mbuf *cookie;
 460
 461         desc = vq->vq_packed.ring.desc;
 462         used_idx = vq->vq_used_cons_idx;
 463         if (!desc_is_used(&desc[used_idx], vq))
 464                 return -1;
 465
 466         len = desc[used_idx].len;
 467         id = desc[used_idx].id;
 468         cookie = (struct rte_mbuf *)vq->vq_descx[id].cookie;
 469         if (unlikely(cookie == NULL)) {
 470                 PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
 471                                 vq->vq_used_cons_idx);
 472                 return -1;
 473         }
 474         rte_prefetch0(cookie);
 475         rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
 476
 477         cookie->data_off = RTE_PKTMBUF_HEADROOM;
 478         cookie->ol_flags = 0;
 479         cookie->pkt_len = (uint32_t)(len - hdr_size);
 480         cookie->data_len = (uint32_t)(len - hdr_size);
 481
 482         hdr = (struct virtio_net_hdr *)((char *)cookie->buf_addr +
 483                                         RTE_PKTMBUF_HEADROOM - hdr_size);
 484         if (hw->has_rx_offload)
 485                 virtio_vec_rx_offload(cookie, hdr);
 486
 487         *rx_pkts = cookie;
 488
 489         rxvq->stats.bytes += cookie->pkt_len;
 490
 491         vq->vq_free_cnt++;
 492         vq->vq_used_cons_idx++;
 493         if (vq->vq_used_cons_idx >= vq->vq_nentries) {
 494                 vq->vq_used_cons_idx -= vq->vq_nentries;
 495                 vq->vq_packed.used_wrap_counter ^= 1;
 496         }
 497
 498         return 0;
 499 }
 500
 501 static inline void
 502 virtio_recv_refill_packed_vec(struct virtnet_rx *rxvq,
 503                               struct rte_mbuf **cookie,
 504                               uint16_t num)
 505 {
 506         struct virtqueue *vq = rxvq->vq;
 507         struct vring_packed_desc *start_dp = vq->vq_packed.ring.desc;
 508         uint16_t flags = vq->vq_packed.cached_flags;
 509         struct virtio_hw *hw = vq->hw;
 510         struct vq_desc_extra *dxp;
 511         uint16_t idx, i;
 512         uint16_t batch_num, total_num = 0;
 513         uint16_t head_idx = vq->vq_avail_idx;
 514         uint16_t head_flag = vq->vq_packed.cached_flags;
 515         uint64_t addr;
 516
 517         do {
 518                 idx = vq->vq_avail_idx;
 519
 520                 batch_num = PACKED_BATCH_SIZE;
 521                 if (unlikely((idx + PACKED_BATCH_SIZE) > vq->vq_nentries))
 522                         batch_num = vq->vq_nentries - idx;
 523                 if (unlikely((total_num + batch_num) > num))
 524                         batch_num = num - total_num;
 525
 526                 virtio_for_each_try_unroll(i, 0, batch_num) {
 527                         dxp = &vq->vq_descx[idx + i];
 528                         dxp->cookie = (void *)cookie[total_num + i];
 529
 530                         addr = VIRTIO_MBUF_ADDR(cookie[total_num + i], vq) +
 531                                 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
 532                         start_dp[idx + i].addr = addr;
 533                         start_dp[idx + i].len = cookie[total_num + i]->buf_len
 534                                 - RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
 535                         if (total_num || i) {
 536                                 virtqueue_store_flags_packed(&start_dp[idx + i],
 537                                                 flags, hw->weak_barriers);
 538                         }
 539                 }
 540
 541                 vq->vq_avail_idx += batch_num;
 542                 if (vq->vq_avail_idx >= vq->vq_nentries) {
 543                         vq->vq_avail_idx -= vq->vq_nentries;
 544                         vq->vq_packed.cached_flags ^=
 545                                 VRING_PACKED_DESC_F_AVAIL_USED;
 546                         flags = vq->vq_packed.cached_flags;
 547                 }
 548                 total_num += batch_num;
 549         } while (total_num < num);
 550
 551         virtqueue_store_flags_packed(&start_dp[head_idx], head_flag,
 552                                 hw->weak_barriers);
 553         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
 554 }
 555
 556 uint16_t
 557 virtio_recv_pkts_packed_vec(void *rx_queue,
 558                             struct rte_mbuf **rx_pkts,
 559                             uint16_t nb_pkts)
 560 {
 561         struct virtnet_rx *rxvq = rx_queue;
 562         struct virtqueue *vq = rxvq->vq;
 563         struct virtio_hw *hw = vq->hw;
 564         uint16_t num, nb_rx = 0;
 565         uint32_t nb_enqueued = 0;
 566         uint16_t free_cnt = vq->vq_free_thresh;
 567
 568         if (unlikely(hw->started == 0))
 569                 return nb_rx;
 570
 571         num = RTE_MIN(VIRTIO_MBUF_BURST_SZ, nb_pkts);
 572         if (likely(num > PACKED_BATCH_SIZE))
 573                 num = num - ((vq->vq_used_cons_idx + num) % PACKED_BATCH_SIZE);
 574
 575         while (num) {
 576                 if (!virtqueue_dequeue_batch_packed_vec(rxvq,
 577                                         &rx_pkts[nb_rx])) {
 578                         nb_rx += PACKED_BATCH_SIZE;
 579                         num -= PACKED_BATCH_SIZE;
 580                         continue;
 581                 }
 582                 if (!virtqueue_dequeue_single_packed_vec(rxvq,
 583                                         &rx_pkts[nb_rx])) {
 584                         nb_rx++;
 585                         num--;
 586                         continue;
 587                 }
 588                 break;
 589         };
 590
 591         PMD_RX_LOG(DEBUG, "dequeue:%d", num);
 592
 593         rxvq->stats.packets += nb_rx;
 594
 595         if (likely(vq->vq_free_cnt >= free_cnt)) {
 596                 struct rte_mbuf *new_pkts[free_cnt];
 597                 if (likely(rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts,
 598                                                 free_cnt) == 0)) {
 599                         virtio_recv_refill_packed_vec(rxvq, new_pkts,
 600                                         free_cnt);
 601                         nb_enqueued += free_cnt;
 602                 } else {
 603                         struct rte_eth_dev *dev =
 604                                 &rte_eth_devices[rxvq->port_id];
 605                         dev->data->rx_mbuf_alloc_failed += free_cnt;
 606                 }
 607         }
 608
 609         if (likely(nb_enqueued)) {
 610                 if (unlikely(virtqueue_kick_prepare_packed(vq))) {
 611                         virtqueue_notify(vq);
 612                         PMD_RX_LOG(DEBUG, "Notified");
 613                 }
 614         }
 615
 616         return nb_rx;
 617 }